<>USD.ai

Princeps is the independent risk and grading layer for compute — call it a credit agency, or even the gold standard, for compute. We price the risk nascent compute markets have so far overlooked, modelling whether a given operator will actually deliver the capacity and uptime it sells. Fundamentally, our thesis is that any GPU-backed loan is only as good as the hardware securing it and the operator running it.

Since last year, compute markets have boomed. Compute now has a price, venues, and a path to physical settlement. What it does not yet have is a deliverable standard (a measure of whose GPU-hours are good delivery and whose are not). Today a top-tier operator and an unreliable one look identical on paper and are procured against the same undifferentiated benchmark. This is more a trust problem than a pricing issue. Historically, every commodity that became tradable first grew an independent layer certifying that what was delivered matched what was promised (gold has the LBMA Good Delivery List, grain has licensed graders and certified warehouses, oil has independent inspection at the delivery point). Liquidity will always follow trust.

We're confident the market is moving in this direction already. Compute is developing observable reference prices through indexes and now futures listings. But a financial contract, as one recent analysis of the compute-offtake market puts it, "cannot make a delayed data centre open on time, secure additional power, repair a network failure, or turn one generation of GPU into another."¹⁷ While a price curve prices the market, it still leaves untouched the basis between a reference price and the compute a specific operator actually delivers (the configuration, delivery, service-level and counterparty risk) that is still bundled and unpriced inside private capacity agreements today. Whoever takes on delivery of that compute, whether the platform reselling it, the lab training on it, or the lender financing it, inherits that risk, not the operator. Princeps prices it. As financing and procurement grow more modular, lenders already separate market-price, utilisation, basis, operating and counterparty risk. Independent of any one index or marketplace, Princeps provides the independent grade for compute based on reliability, capacity and delivery.

offers

• NeoCloud Credit Score — grading of ~940 neocloud operators, from satellite imagery of on-site generation, permits, chip inventory and specs, grid-interconnection age, probabilistic weather and external order data. None of them file anything, so this is the only underwriting file that exists on them. Sold today to compute marketplaces, AI labs and lenders financing GPU fleets, who otherwise have no underwriting file on a private operator that files nothing publicly.
• Verified tier — the top grade, corroborated by an observed delivery record: spin-up success, advertised-vs-actual availability, and completion against real orders. No probes inside an operator's stack, no device install.
• Risk engine — a continuous, per-contract quantification of expected and tail delivery loss, and the foundation of an independent compute credit rating. Initially developed through Princeps data partnership with Ornn.
• On-site hardware telemetry — through our hardware partner ArcLeap, on-premise telemetry installed directly at the site, giving verified serial-level presence, liveness and utilisation on the actual pledged hardware. Live today across 30+ tribal data-centre sites, with new sites in the pipeline currently applying for financing.
• Insurance capacity — Princeps has filed to operate as a Managing General Agent (MGA) in Texas, the basis for underwriting and placing compute-risk cover across the US. This turns the grade into actual coverage — the path by which the risk we quantify becomes insurable, not just priced.

Founding Team

Anya Trofimova

Oxon.

Conrad Frøyland Moe

Oxon.

Wilhem Hector

MIT Mech Eng. & Oxon.

Backed by Y Combinator and angels and advisors from Standard Intelligence, SF Compute, Kojo, Palantir and 8VC.

· Scope of proposed collaboration with USD.ai

USD.ai has built fast, GPU-backed credit that clears in days against hardware tokenised into on-chain warehouse receipts under CALIBER. What is yet to be built is an independent, continuous verification layer: proof that the collateral is physically present, serialised and live, and an independent grade of whether the operator behind it will keep generating the uptime that services the loan. Princeps proposes exactly that layer, in partnership with USD.ai.

Why does this matter for USD.ai? USD.ai originates credit against GPUs held under a datacenter bailment and tokenised into warehouse receipts (GWRTs). Risk is assessed at origination but left to drift post-financing. Princeps solves two questions. First, does the collateral actually exist and stay live on-site — a borrower can move, resell, double-pledge or power down hardware between inspections. Second, will the operator keep delivering — reliability can decay long before a missed payment reveals it. In a default, the remedy is physical repossession of hardware you have never independently verified is there. That risk ultimately lands on sUSDai depositors, who take it today largely on trust.

What we are proposing? One integration that provides both continuous proof-of-collateral and an independent operator grade. Every GWRT gets a live attestation that its serialised hardware is present and online; every borrower gets a Collateral & Credit Grade that maps to an LTV / haircut and an early-warning signal on SLA breach. All of this is auditable, so sUSDai depositors and USD.ai's own investors can verify that every liability is backed by live hardware.

How do we grade? The grade composes two factor groups: the site's physical resilience (from the Princeps database) and the operator's observed collateral & delivery record (from the bailment, loan-servicing history and on-site attestation) into a single score.

Product Preview

Princeps already grades neocloud facilities in production, from reviewed engineering and public-market evidence, under a governed, versioned methodology. Below is the live workspace applied to operators across a cross-section of London facilities.

01 What makes collateral verification as meaningful as yield?

A GWRT records that a borrower pledged a specific fleet of GPUs, held on-site under a bailment. What it does not record, after issuance, is whether those exact serial numbers are still racked, still powered, and still generating the revenue that services the loan. The underwriting file at origination is a snapshot, while collateral progressively deteriorates and performs unreliably. USD.ai originates credit only against hardware that is "active, on-site, and operating under a valid datacenter bailment"⁶, but the open question of reliability still remains.

Ownership on paper is not hardware on-site. A GWRT records who owns the GPUs, not whether they are still racked and running — so on-chain, a depleted fleet looks identical to a full one.

What separates a good operator from a bad one is not whether hardware fails (which it inevitably will) but how well that run is contained. In one study, Meta held >90% effective training time because only 3 of 419 incidents needed manual intervention.⁷ A weaker operator can run up to 10.7% lost compute. The failure waterfall scales with cluster size, meaning that the larger the reserve, the more a run depends on the operator's containment and rerouting capabilities.

02 An absence of existing market signals

Princeps reconstructs the loss history for neoclouds and compute providers from zero to one. Our fundamental belief is that risk must be assessed at origination, across both the physical and performance layers, to be meaningfully quantified. This is the single most mispriced risk in GPU-backed lending right now. Uptime Institute tiers measure redundancy, not AI-readiness (built for 5–10 kW racks when a GB200 rack exceeds 120 kW) and certification is static.¹⁰ SLAs cap the remedy at a service credit, usually a 10% credit on a ~$3 instance is about 30 cents, against outages averaging near $1 million.¹¹ The most serious independent effort is ClusterMAX. SemiAnalysis is both the strongest proof that reliability varies enormously and that an independent grade has real traction.

ClusterMAX is a periodic, editorial tier list refreshed quarterly. Princeps extends the same evidence base into a continuous, per-loan risk quantification a lender can act on in real time as a live benchmark. We are independent of any one index or marketplace, allowing us to collect data across every operator, site and venue. Princeps fills this intelligence for both the operator and the lender financing it.

03 Methodology

Reliability can be inferred from the physical and structural signals around a provider, and confirmed by whether the provider actually delivered against real orders. Princeps grades ~940 operators into five bands. Crucially, the baseline grade requires no probes inside an operator's stack. It is built from external data plus the bailment and loan-servicing record, and is deepened, at the operator's option, by device-level attestation of the pledged serials.

04 The loan-loss benchmark

Every grade resolves to a dollar figure a credit team can underwrite against. Dial in the facility (fleet, size, term, checkpoint interval) and Princeps runs the Monte-Carlo model to translate the operator's reliability into expected and tail loss on that specific loan.

Sample demonstration:

05 Failure simulation

The loss figures are derived from a Monte-Carlo simulation of the facility you set above. Each of 4,000 trials draws on both hardware failures⁹ and a severity per failure (how much revenue-generating uptime is lost and how long recovery takes, scaled by the operator's containment — the grade). The result is a full distribution of serviced value. The gap between the expected outcome and the unlucky tail (P99) is the loss a lender carries silently today, and exactly what a collateral & credit grade lets USD.ai surface, quantify and price into LTV.

Model: failures ~ Poisson(N·24·T / 50,000); per-failure lost cluster-time ~ checkpoint-interval rework + recovery, exponential severity, containment scaled so each grade's mean matches its measured goodput-loss band [SemiAnalysis/Nebius]. Illustrative — for structure, not a quote. Synchronous training assumed (a failure stalls the whole cluster to the last checkpoint).

The grade applied to a representative GPU-backed collateral pool

We joined six operators representative of a GPU-backed loan pool to the Princeps database and pulled each one's independent tier, ownership model, chip generation and site footprint.

And in the lender's dashboard, it looks like this

· Modelling LTV, drift, portfolio loss, cost of capital

A grade fundamentally sets out how much can be safely advanced against a given fleet, because a stronger operator's collateral is worth more against the same hardware. It catches collateral that is quietly deteriorating long before a payment is missed and sizes how much loss the book is actually carrying, so depositors can see the buffer that protects them. Princeps also separates the true cost of capital from the uncertainty premium a lender pays simply for not knowing.

Grounded in USD.ai's own figures, on an identical H100 fleet, the operator grade moves the defensible advance rate from the low-40s to USD.ai's full 70% tier — a swing of roughly 25 points in how much capital the same collateral safely supports, purely on who is running it. It moves modelled annual loss by close to an order of magnitude, from a few tenths of a percent under a Verified operator to several percent under an Elevated one. Carried across the $205M deployed book, grading that spread through the haircut roughly halves expected loss and materially thins the tail. On the pricing side, most of a ~15% borrower rate over the risk-free base is not compensation for credit risk at all, it is an uncertainty premium paid for underwriting blind. An independent grade converts that premium directly into cheaper capital for reliable operators and wider, more durable margin for the protocol.

Grade → advance rate, and drift between inspections

Depositor protection and the cost of capital

Illustrative models for structure, not quotes. LTV and expected loss are functions of the operator grade; portfolio figures apply grade-driven haircuts across a representative pool; the cost-of-capital view decomposes a typical GPU-loan APR into a base rate and the uncertainty premium an independent grade compresses — the buffer, and the margin, that stand between the loan book and sUSDai depositors.

06 The impact of an independent verification & credit layer

That a grade transforms a credit market is one of the better-established results in economics. Akerlof's 1970 "market for lemons" showed that when buyers can't observe quality they pay only for average quality, good sellers exit, and the market can unravel. The fix is the creation of "counteracting institutions" such as certification and third-party grading.⁴

The mechanism is identical: an independent layer converts an unobservable quality attribute into a verifiable, priceable signal — raising prices for good sellers, raising transaction probability, widening participation, and shrinking the pooling that drives adverse selection.

The cost of capital falls for good operators. A grade is most valuable exactly where trust is scarcest and borrowers are otherwise indistinguishable, as in GPU-backed lending. Capital shifts to quality, and the depositor pool widens: allocators commit to sUSDai on verifiable, auditable backing, not trust. An independent, quantified collateral & credit grade brings in capital that is currently sitting out for want of exactly this assurance.

07 Scope of proposed collaboration

• Proof-of-collateral attestation — a Princeps API that verifies each GWRT's pledged serial numbers are present, powered and reachable on-site, and flags any that go dark, move or get re-pledged. Point-in-time inspection becomes a continuous, auditable feed.
• A Collateral & Credit Grade → LTV map — every borrower graded on physical resilience plus collateral & delivery record; the grade sets the haircut, so stronger operators borrow more and weaker ones are priced or gated out.
• An SLA-breach early-warning feed — the live loss benchmark exposed via API: for any facility, a quantified expected and tail loss and a breach signal that fires well before a missed payment, so USD.ai and its curators can act early.
• Enforceable remedy on default — through our hardware partner ArcLeap, the same attestation channel supports remote verification and, where contractually permitted, remote disablement of the pledged hardware — turning "complex, delayed repossession" into an executable control. USD.ai's loan-servicing record in turn sharpens our grade through a closed data loop.

How Princeps completes USD.ai underwriting and risk modelling

• Independent underwriting. Today independent experts vet every borrower and asset so only high-quality hardware backs the protocol — a bespoke, manual process.⁶ A standardised Princeps operator grade becomes the underwriting input that decides who qualifies and at what LTV, turning "we trust these reviewers" into "here is the score."
• Collateral eligibility & rate tiers. CHIP governance already sets which hardware qualifies as collateral and how interest-rate tiers are structured across asset classes. Our grading — across chip generation, grid interconnection, permits and chip inventory — maps directly onto both: higher-grade operators earn better rate tiers and higher LTV.
• Ongoing covenant monitoring. These are non-recourse loans, so the protocol absorbs the loss if collateral stops producing. Device-level telemetry through ArcLeap is a continuous covenant-monitoring and early-warning feed on deteriorating operators — protecting the loan book and, by extension, sUSDai holders.
• Pricing the risk premium. USD.ai is explicit that capital locked until repayment demands a higher return. A trusted, independent grade shrinks that uncertainty premium — a direct, quantifiable improvement to the protocol's cost of capital, and the cleanest place to start the conversation.

The Princeps intelligence layer

Underneath the grade sits a continuous intelligence layer that scores every operator across six dimensions:

Princeps would provide ongoing risk intelligence, collateral verification and credit scoring for USD.ai — a live index across every operator in the loan book. It can run as a standalone underwriting-and-audit service for USD.ai and its curators, or integrate directly into the protocol: grade-driven LTV, the attestation feed, and an investor-facing audit view all drawing on the same index.

Joining the Princeps layer to the USD.ai ecosystem

The ecosystem already has a partner for every layer except this one. We would slot in as the independent Risk & Collateral-Verification category — the missing counterpart to the data and security partners USD.ai already relies on.

The End Goal: the collateral & credit standard for GPU-backed lending

The integration above is the starting point, and we are keen to explore a larger, longer-term collaboration alongside it. USD.ai generates data most of the market cannot see — a continuous record of which operators actually serviced their loans and whose collateral stayed live, drawn from every facility it originates. Combined with the Princeps methodology, that record is the foundation for a jointly-operated Compute Credit Rating. Over time it would position USD.ai as the reference point for GPU-backed credit, much as the Chicago Board of Trade's grading standards became the basis on which the grain market settled.⁵

The path from independent grade to rating standard runs in three phases:

• Now — Collateral & Credit Grade, proof-of-collateral attestation and grade-driven LTV, with no data dependency. USD.ai gets an independent verification and underwriting layer.
• Next — a Verified tier plus continuous, calibrated loss scoring, corroborated by USD.ai's own loan-servicing data and an investor-facing audit feed.
• Compute credit standard — a published, continuously-updated Compute Credit Rating, with USD.ai's origination and servicing data calibrated to our reconstructed loss history. Together it becomes the reference rating the whole GPU-backed credit market prices off, lowering the cost of capital for reliable operators.

Princeps operates the rating independently, with USD.ai as the data spine and reference venue. A rating the lender controls is not an independent rating. Independence is what makes the standard credible to depositors, investors and counterparties and it is what underpins protocol solvency and depositor confidence over the long term. The rating improves the more volume USD.ai originates.